[Paper Review] Data consistency networks for (calibration-less) accelerated parallel MR image reconstruction

Jo Schlemper, Jinming Duan+5 more|arXiv (Cornell University)|Sep 25, 2019

Medical Imaging Techniques and Applications5 citations

TL;DR

This paper proposes two deep learning-based MRI reconstruction networks—D-POCSENSE and DC-CNN—that extend the deep cascade of CNNs by integrating data consistency layers for accelerated parallel MRI reconstruction. D-POCSENSE uses coil sensitivity maps for improved image quality, while DC-CNN is calibration-less and processes multi-coil data jointly; both achieve state-of-the-art performance in PSNR and SSIM on a public knee dataset with acceleration factors of 4 and 6.

ABSTRACT

We present simple reconstruction networks for multi-coil data by extending deep cascade of CNN's and exploiting the data consistency layer. In particular, we propose two variants, where one is inspired by POCSENSE and the other is calibration-less. We show that the proposed approaches are competitive relative to the state of the art both quantitatively and qualitatively.

Motivation & Objective

• Address the challenge of accelerating parallel MRI reconstruction while avoiding calibration-dependent methods.
• Improve image quality in highly undersampled k-space data using deep learning with data consistency constraints.
• Develop a calibration-less alternative to existing sensitivity-based methods like POCSENSE for greater clinical practicality.
• Achieve competitive performance compared to state-of-the-art methods such as Variational Network and ℓ₁-SPIRiT.
• Demonstrate generalization across multiple acquisition protocols (T2, PD, fat-sat) and acceleration factors (4, 6).

Proposed method

• Extend the deep cascade of CNNs (DC-CNN) by interleaving denoising sub-networks with data consistency layers for k-space updates.
• Propose D-POCSENSE: uses sensitivity-weighted recombined image input and applies data consistency via coil-wise k-space update with a trainable λ balancing term.
• Implement DC-CNN: processes all coil images jointly by stacking them along the channel axis and applying data consistency independently per coil.
• Use a trainable λ in the data consistency layer to adapt to noise levels, improving robustness.
• Train both networks using ℓ₂ loss (D-POCSENSE) and weighted ℓ₂ loss (DC-CNN) with Adam optimizer over 200 epochs.
• Apply k-space sampling with 24 central lines as calibration region and Cartesian undersampling at AF=4 and 6.

Experimental results

Research questions

• RQ1 Can a deep cascade with data consistency layers outperform existing state-of-the-art methods in accelerated parallel MRI reconstruction?
• RQ2 Does incorporating coil sensitivity maps in the network architecture (D-POCSENSE) lead to better reconstruction quality than a calibration-less approach (DC-CNN)?
• RQ3 Can a calibration-less method (DC-CNN) achieve competitive performance without requiring sensitivity maps or calibration data?
• RQ4 How do the proposed methods compare in terms of PSNR and SSIM across diverse MRI protocols and acceleration factors?
• RQ5 Does longer training improve the performance of D-POCSENSE and DC-CNN, especially in reducing residual aliasing?

Key findings

• D-POCSENSE achieved a PSNR of 36.03 ± 10.26 and SSIM of 0.90 ± 0.24 at AF=4 on axial T2 FS, outperforming ℓ₁-SPIRiT and matching the performance of the Variational Network.
• At AF=6, D-POCSENSE achieved PSNR of 31.83 ± 8.79 and SSIM of 0.87 ± 0.25 on axial T2 FS, outperforming ℓ₁-SPIRiT and showing comparable results to VN.
• DC-CNN achieved a PSNR of 35.45 ± 8.63 and SSIM of 0.91 ± 0.19 at AF=4 on axial T2 FS, demonstrating strong performance despite being calibration-less.
• For coronal PD, D-POCSENSE achieved PSNR of 36.94 ± 1.24 and SSIM of 0.98 ± 0.00 at AF=4, matching the performance of the Variational Network.
• At AF=6, the Variational Network achieved the highest PSNR (32.39 ± 9.94) on axial T2 FS, but D-POCSENSE and DC-CNN remained competitive with lower variance.
• Visual results showed D-POCSENSE produced the most homogeneous images with less aliasing, while DC-CNN and VN showed residual artifacts at AF=6.